A study on high-order hidden Markov models and applications to speech recognition

Abstract

We propose high-order hidden Markov models (HO-HMM) to capture the duration and dynamics of speech signal, In the proposed model, both the state transition probability and the output observation probability depend not only on the current state but also on several previous states. An extended Viterbi algorithm was developed to train model and recognize speech. The performance of the HO-HMM was investigated by conducting experiments on speaker independent Mandarin digits recognition. From the experimental results, we find that as the order of HO-HMM increases, the number of error reduces. We also find that systems with both high-order state transition probability distribution and output observation probability distribution outperform systems with only high-order state transition probability distribution. © Springer-Verlag Berlin Heidelberg 2006.